1 /* 2 * Copyright (c) 1997, 2020, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "jvm.h" 27 #include "classfile/javaClasses.inline.hpp" 28 #include "classfile/symbolTable.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/codeCache.hpp" 31 #include "code/debugInfoRec.hpp" 32 #include "code/nmethod.hpp" 33 #include "code/pcDesc.hpp" 34 #include "code/scopeDesc.hpp" 35 #include "compiler/compilationPolicy.hpp" 36 #include "interpreter/bytecode.hpp" 37 #include "interpreter/interpreter.hpp" 38 #include "interpreter/oopMapCache.hpp" 39 #include "memory/allocation.inline.hpp" 40 #include "memory/oopFactory.hpp" 41 #include "memory/resourceArea.hpp" 42 #include "memory/universe.hpp" 43 #include "oops/constantPool.hpp" 44 #include "oops/method.hpp" 45 #include "oops/objArrayKlass.hpp" 46 #include "oops/objArrayOop.inline.hpp" 47 #include "oops/oop.inline.hpp" 48 #include "oops/fieldStreams.inline.hpp" 49 #include "oops/typeArrayOop.inline.hpp" 50 #include "oops/verifyOopClosure.hpp" 51 #include "prims/jvmtiThreadState.hpp" 52 #include "runtime/atomic.hpp" 53 #include "runtime/biasedLocking.hpp" 54 #include "runtime/deoptimization.hpp" 55 #include "runtime/fieldDescriptor.hpp" 56 #include "runtime/fieldDescriptor.inline.hpp" 57 #include "runtime/frame.inline.hpp" 58 #include "runtime/handles.inline.hpp" 59 #include "runtime/interfaceSupport.inline.hpp" 60 #include "runtime/jniHandles.inline.hpp" 61 #include "runtime/safepointVerifiers.hpp" 62 #include "runtime/sharedRuntime.hpp" 63 #include "runtime/signature.hpp" 64 #include "runtime/stubRoutines.hpp" 65 #include "runtime/thread.hpp" 66 #include "runtime/threadSMR.hpp" 67 #include "runtime/vframe.hpp" 68 #include "runtime/vframeArray.hpp" 69 #include "runtime/vframe_hp.hpp" 70 #include "utilities/events.hpp" 71 #include "utilities/macros.hpp" 72 #include "utilities/preserveException.hpp" 73 #include "utilities/xmlstream.hpp" 74 #if INCLUDE_JFR 75 #include "jfr/jfrEvents.hpp" 76 #include "jfr/metadata/jfrSerializer.hpp" 77 #endif 78 79 bool DeoptimizationMarker::_is_active = false; 80 81 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 82 int caller_adjustment, 83 int caller_actual_parameters, 84 int number_of_frames, 85 intptr_t* frame_sizes, 86 address* frame_pcs, 87 BasicType return_type, 88 int exec_mode) { 89 _size_of_deoptimized_frame = size_of_deoptimized_frame; 90 _caller_adjustment = caller_adjustment; 91 _caller_actual_parameters = caller_actual_parameters; 92 _number_of_frames = number_of_frames; 93 _frame_sizes = frame_sizes; 94 _frame_pcs = frame_pcs; 95 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 96 _return_type = return_type; 97 _initial_info = 0; 98 // PD (x86 only) 99 _counter_temp = 0; 100 _unpack_kind = exec_mode; 101 _sender_sp_temp = 0; 102 103 _total_frame_sizes = size_of_frames(); 104 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode"); 105 } 106 107 108 Deoptimization::UnrollBlock::~UnrollBlock() { 109 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 110 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 111 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 112 } 113 114 115 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 116 assert(register_number < RegisterMap::reg_count, "checking register number"); 117 return &_register_block[register_number * 2]; 118 } 119 120 121 122 int Deoptimization::UnrollBlock::size_of_frames() const { 123 // Acount first for the adjustment of the initial frame 124 int result = _caller_adjustment; 125 for (int index = 0; index < number_of_frames(); index++) { 126 result += frame_sizes()[index]; 127 } 128 return result; 129 } 130 131 132 void Deoptimization::UnrollBlock::print() { 133 ttyLocker ttyl; 134 tty->print_cr("UnrollBlock"); 135 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 136 tty->print( " frame_sizes: "); 137 for (int index = 0; index < number_of_frames(); index++) { 138 tty->print(INTX_FORMAT " ", frame_sizes()[index]); 139 } 140 tty->cr(); 141 } 142 143 144 // In order to make fetch_unroll_info work properly with escape 145 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 146 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 147 // of previously eliminated objects occurs in realloc_objects, which is 148 // called from the method fetch_unroll_info_helper below. 149 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode)) 150 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 151 // but makes the entry a little slower. There is however a little dance we have to 152 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 153 154 // fetch_unroll_info() is called at the beginning of the deoptimization 155 // handler. Note this fact before we start generating temporary frames 156 // that can confuse an asynchronous stack walker. This counter is 157 // decremented at the end of unpack_frames(). 158 if (TraceDeoptimization) { 159 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread)); 160 } 161 thread->inc_in_deopt_handler(); 162 163 return fetch_unroll_info_helper(thread, exec_mode); 164 JRT_END 165 166 #if COMPILER2_OR_JVMCI 167 static bool eliminate_allocations(JavaThread* thread, int exec_mode, CompiledMethod* compiled_method, 168 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk) { 169 bool realloc_failures = false; 170 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 171 172 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 173 174 // The flag return_oop() indicates call sites which return oop 175 // in compiled code. Such sites include java method calls, 176 // runtime calls (for example, used to allocate new objects/arrays 177 // on slow code path) and any other calls generated in compiled code. 178 // It is not guaranteed that we can get such information here only 179 // by analyzing bytecode in deoptimized frames. This is why this flag 180 // is set during method compilation (see Compile::Process_OopMap_Node()). 181 // If the previous frame was popped or if we are dispatching an exception, 182 // we don't have an oop result. 183 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt); 184 Handle return_value; 185 if (save_oop_result) { 186 // Reallocation may trigger GC. If deoptimization happened on return from 187 // call which returns oop we need to save it since it is not in oopmap. 188 oop result = deoptee.saved_oop_result(&map); 189 assert(oopDesc::is_oop_or_null(result), "must be oop"); 190 return_value = Handle(thread, result); 191 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 192 if (TraceDeoptimization) { 193 ttyLocker ttyl; 194 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); 195 } 196 } 197 if (objects != NULL) { 198 JRT_BLOCK 199 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD); 200 JRT_END 201 bool skip_internal = (compiled_method != NULL) && !compiled_method->is_compiled_by_jvmci(); 202 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal); 203 #ifndef PRODUCT 204 if (TraceDeoptimization) { 205 ttyLocker ttyl; 206 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 207 Deoptimization::print_objects(objects, realloc_failures); 208 } 209 #endif 210 } 211 if (save_oop_result) { 212 // Restore result. 213 deoptee.set_saved_oop_result(&map, return_value()); 214 } 215 return realloc_failures; 216 } 217 218 static void eliminate_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 219 HandleMark hm; 220 #ifndef PRODUCT 221 bool first = true; 222 #endif 223 for (int i = 0; i < chunk->length(); i++) { 224 compiledVFrame* cvf = chunk->at(i); 225 assert (cvf->scope() != NULL,"expect only compiled java frames"); 226 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 227 if (monitors->is_nonempty()) { 228 Deoptimization::relock_objects(monitors, thread, realloc_failures); 229 #ifndef PRODUCT 230 if (PrintDeoptimizationDetails) { 231 ttyLocker ttyl; 232 for (int j = 0; j < monitors->length(); j++) { 233 MonitorInfo* mi = monitors->at(j); 234 if (mi->eliminated()) { 235 if (first) { 236 first = false; 237 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 238 } 239 if (mi->owner_is_scalar_replaced()) { 240 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 241 tty->print_cr(" failed reallocation for klass %s", k->external_name()); 242 } else { 243 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner())); 244 } 245 } 246 } 247 } 248 #endif // !PRODUCT 249 } 250 } 251 } 252 #endif // COMPILER2_OR_JVMCI 253 254 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 255 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) { 256 257 // Note: there is a safepoint safety issue here. No matter whether we enter 258 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 259 // the vframeArray is created. 260 // 261 262 // Allocate our special deoptimization ResourceMark 263 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 264 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 265 thread->set_deopt_mark(dmark); 266 267 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 268 RegisterMap map(thread, true); 269 RegisterMap dummy_map(thread, false); 270 // Now get the deoptee with a valid map 271 frame deoptee = stub_frame.sender(&map); 272 // Set the deoptee nmethod 273 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!"); 274 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); 275 thread->set_deopt_compiled_method(cm); 276 277 if (VerifyStack) { 278 thread->validate_frame_layout(); 279 } 280 281 // Create a growable array of VFrames where each VFrame represents an inlined 282 // Java frame. This storage is allocated with the usual system arena. 283 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 284 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 285 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 286 while (!vf->is_top()) { 287 assert(vf->is_compiled_frame(), "Wrong frame type"); 288 chunk->push(compiledVFrame::cast(vf)); 289 vf = vf->sender(); 290 } 291 assert(vf->is_compiled_frame(), "Wrong frame type"); 292 chunk->push(compiledVFrame::cast(vf)); 293 294 bool realloc_failures = false; 295 296 #if COMPILER2_OR_JVMCI 297 #if INCLUDE_JVMCI 298 bool jvmci_enabled = true; 299 #else 300 bool jvmci_enabled = false; 301 #endif 302 303 // Reallocate the non-escaping objects and restore their fields. Then 304 // relock objects if synchronization on them was eliminated. 305 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations) )) { 306 realloc_failures = eliminate_allocations(thread, exec_mode, cm, deoptee, map, chunk); 307 } 308 #endif // COMPILER2_OR_JVMCI 309 310 // Revoke biases, done with in java state. 311 // No safepoints allowed after this 312 revoke_from_deopt_handler(thread, deoptee, &map); 313 314 // Ensure that no safepoint is taken after pointers have been stored 315 // in fields of rematerialized objects. If a safepoint occurs from here on 316 // out the java state residing in the vframeArray will be missed. 317 // Locks may be rebaised in a safepoint. 318 NoSafepointVerifier no_safepoint; 319 320 #if COMPILER2_OR_JVMCI 321 if (jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) )) { 322 eliminate_locks(thread, chunk, realloc_failures); 323 } 324 #endif // COMPILER2_OR_JVMCI 325 326 ScopeDesc* trap_scope = chunk->at(0)->scope(); 327 Handle exceptionObject; 328 if (trap_scope->rethrow_exception()) { 329 if (PrintDeoptimizationDetails) { 330 tty->print_cr("Exception to be rethrown in the interpreter for method %s::%s at bci %d", trap_scope->method()->method_holder()->name()->as_C_string(), trap_scope->method()->name()->as_C_string(), trap_scope->bci()); 331 } 332 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 333 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw"); 334 ScopeValue* topOfStack = expressions->top(); 335 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 336 guarantee(exceptionObject() != NULL, "exception oop can not be null"); 337 } 338 339 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures); 340 #if COMPILER2_OR_JVMCI 341 if (realloc_failures) { 342 pop_frames_failed_reallocs(thread, array); 343 } 344 #endif 345 346 assert(thread->vframe_array_head() == NULL, "Pending deopt!"); 347 thread->set_vframe_array_head(array); 348 349 // Now that the vframeArray has been created if we have any deferred local writes 350 // added by jvmti then we can free up that structure as the data is now in the 351 // vframeArray 352 353 if (thread->deferred_locals() != NULL) { 354 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 355 int i = 0; 356 do { 357 // Because of inlining we could have multiple vframes for a single frame 358 // and several of the vframes could have deferred writes. Find them all. 359 if (list->at(i)->id() == array->original().id()) { 360 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 361 list->remove_at(i); 362 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 363 delete dlv; 364 } else { 365 i++; 366 } 367 } while ( i < list->length() ); 368 if (list->length() == 0) { 369 thread->set_deferred_locals(NULL); 370 // free the list and elements back to C heap. 371 delete list; 372 } 373 374 } 375 376 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 377 CodeBlob* cb = stub_frame.cb(); 378 // Verify we have the right vframeArray 379 assert(cb->frame_size() >= 0, "Unexpected frame size"); 380 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 381 382 // If the deopt call site is a MethodHandle invoke call site we have 383 // to adjust the unpack_sp. 384 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 385 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 386 unpack_sp = deoptee.unextended_sp(); 387 388 #ifdef ASSERT 389 assert(cb->is_deoptimization_stub() || 390 cb->is_uncommon_trap_stub() || 391 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 392 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 393 "unexpected code blob: %s", cb->name()); 394 #endif 395 396 // This is a guarantee instead of an assert because if vframe doesn't match 397 // we will unpack the wrong deoptimized frame and wind up in strange places 398 // where it will be very difficult to figure out what went wrong. Better 399 // to die an early death here than some very obscure death later when the 400 // trail is cold. 401 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 402 // in that it will fail to detect a problem when there is one. This needs 403 // more work in tiger timeframe. 404 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 405 406 int number_of_frames = array->frames(); 407 408 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 409 // virtual activation, which is the reverse of the elements in the vframes array. 410 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 411 // +1 because we always have an interpreter return address for the final slot. 412 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 413 int popframe_extra_args = 0; 414 // Create an interpreter return address for the stub to use as its return 415 // address so the skeletal frames are perfectly walkable 416 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 417 418 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 419 // activation be put back on the expression stack of the caller for reexecution 420 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 421 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 422 } 423 424 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 425 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 426 // than simply use array->sender.pc(). This requires us to walk the current set of frames 427 // 428 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 429 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 430 431 // It's possible that the number of parameters at the call site is 432 // different than number of arguments in the callee when method 433 // handles are used. If the caller is interpreted get the real 434 // value so that the proper amount of space can be added to it's 435 // frame. 436 bool caller_was_method_handle = false; 437 if (deopt_sender.is_interpreted_frame()) { 438 methodHandle method(thread, deopt_sender.interpreter_frame_method()); 439 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 440 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 441 // Method handle invokes may involve fairly arbitrary chains of 442 // calls so it's impossible to know how much actual space the 443 // caller has for locals. 444 caller_was_method_handle = true; 445 } 446 } 447 448 // 449 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 450 // frame_sizes/frame_pcs[1] next oldest frame (int) 451 // frame_sizes/frame_pcs[n] youngest frame (int) 452 // 453 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 454 // owns the space for the return address to it's caller). Confusing ain't it. 455 // 456 // The vframe array can address vframes with indices running from 457 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 458 // When we create the skeletal frames we need the oldest frame to be in the zero slot 459 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 460 // so things look a little strange in this loop. 461 // 462 int callee_parameters = 0; 463 int callee_locals = 0; 464 for (int index = 0; index < array->frames(); index++ ) { 465 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 466 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 467 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 468 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 469 callee_locals, 470 index == 0, 471 popframe_extra_args); 472 // This pc doesn't have to be perfect just good enough to identify the frame 473 // as interpreted so the skeleton frame will be walkable 474 // The correct pc will be set when the skeleton frame is completely filled out 475 // The final pc we store in the loop is wrong and will be overwritten below 476 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 477 478 callee_parameters = array->element(index)->method()->size_of_parameters(); 479 callee_locals = array->element(index)->method()->max_locals(); 480 popframe_extra_args = 0; 481 } 482 483 // Compute whether the root vframe returns a float or double value. 484 BasicType return_type; 485 { 486 methodHandle method(thread, array->element(0)->method()); 487 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 488 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 489 } 490 491 // Compute information for handling adapters and adjusting the frame size of the caller. 492 int caller_adjustment = 0; 493 494 // Compute the amount the oldest interpreter frame will have to adjust 495 // its caller's stack by. If the caller is a compiled frame then 496 // we pretend that the callee has no parameters so that the 497 // extension counts for the full amount of locals and not just 498 // locals-parms. This is because without a c2i adapter the parm 499 // area as created by the compiled frame will not be usable by 500 // the interpreter. (Depending on the calling convention there 501 // may not even be enough space). 502 503 // QQQ I'd rather see this pushed down into last_frame_adjust 504 // and have it take the sender (aka caller). 505 506 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 507 caller_adjustment = last_frame_adjust(0, callee_locals); 508 } else if (callee_locals > callee_parameters) { 509 // The caller frame may need extending to accommodate 510 // non-parameter locals of the first unpacked interpreted frame. 511 // Compute that adjustment. 512 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 513 } 514 515 // If the sender is deoptimized the we must retrieve the address of the handler 516 // since the frame will "magically" show the original pc before the deopt 517 // and we'd undo the deopt. 518 519 frame_pcs[0] = deopt_sender.raw_pc(); 520 521 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 522 523 #if INCLUDE_JVMCI 524 if (exceptionObject() != NULL) { 525 thread->set_exception_oop(exceptionObject()); 526 exec_mode = Unpack_exception; 527 } 528 #endif 529 530 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) { 531 assert(thread->has_pending_exception(), "should have thrown OOME"); 532 thread->set_exception_oop(thread->pending_exception()); 533 thread->clear_pending_exception(); 534 exec_mode = Unpack_exception; 535 } 536 537 #if INCLUDE_JVMCI 538 if (thread->frames_to_pop_failed_realloc() > 0) { 539 thread->set_pending_monitorenter(false); 540 } 541 #endif 542 543 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 544 caller_adjustment * BytesPerWord, 545 caller_was_method_handle ? 0 : callee_parameters, 546 number_of_frames, 547 frame_sizes, 548 frame_pcs, 549 return_type, 550 exec_mode); 551 // On some platforms, we need a way to pass some platform dependent 552 // information to the unpacking code so the skeletal frames come out 553 // correct (initial fp value, unextended sp, ...) 554 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 555 556 if (array->frames() > 1) { 557 if (VerifyStack && TraceDeoptimization) { 558 ttyLocker ttyl; 559 tty->print_cr("Deoptimizing method containing inlining"); 560 } 561 } 562 563 array->set_unroll_block(info); 564 return info; 565 } 566 567 // Called to cleanup deoptimization data structures in normal case 568 // after unpacking to stack and when stack overflow error occurs 569 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 570 vframeArray *array) { 571 572 // Get array if coming from exception 573 if (array == NULL) { 574 array = thread->vframe_array_head(); 575 } 576 thread->set_vframe_array_head(NULL); 577 578 // Free the previous UnrollBlock 579 vframeArray* old_array = thread->vframe_array_last(); 580 thread->set_vframe_array_last(array); 581 582 if (old_array != NULL) { 583 UnrollBlock* old_info = old_array->unroll_block(); 584 old_array->set_unroll_block(NULL); 585 delete old_info; 586 delete old_array; 587 } 588 589 // Deallocate any resource creating in this routine and any ResourceObjs allocated 590 // inside the vframeArray (StackValueCollections) 591 592 delete thread->deopt_mark(); 593 thread->set_deopt_mark(NULL); 594 thread->set_deopt_compiled_method(NULL); 595 596 597 if (JvmtiExport::can_pop_frame()) { 598 #ifndef CC_INTERP 599 // Regardless of whether we entered this routine with the pending 600 // popframe condition bit set, we should always clear it now 601 thread->clear_popframe_condition(); 602 #else 603 // C++ interpreter will clear has_pending_popframe when it enters 604 // with method_resume. For deopt_resume2 we clear it now. 605 if (thread->popframe_forcing_deopt_reexecution()) 606 thread->clear_popframe_condition(); 607 #endif /* CC_INTERP */ 608 } 609 610 // unpack_frames() is called at the end of the deoptimization handler 611 // and (in C2) at the end of the uncommon trap handler. Note this fact 612 // so that an asynchronous stack walker can work again. This counter is 613 // incremented at the beginning of fetch_unroll_info() and (in C2) at 614 // the beginning of uncommon_trap(). 615 thread->dec_in_deopt_handler(); 616 } 617 618 // Moved from cpu directories because none of the cpus has callee save values. 619 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 620 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 621 622 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 623 // the days we had adapter frames. When we deoptimize a situation where a 624 // compiled caller calls a compiled caller will have registers it expects 625 // to survive the call to the callee. If we deoptimize the callee the only 626 // way we can restore these registers is to have the oldest interpreter 627 // frame that we create restore these values. That is what this routine 628 // will accomplish. 629 630 // At the moment we have modified c2 to not have any callee save registers 631 // so this problem does not exist and this routine is just a place holder. 632 633 assert(f->is_interpreted_frame(), "must be interpreted"); 634 } 635 636 // Return BasicType of value being returned 637 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 638 639 // We are already active in the special DeoptResourceMark any ResourceObj's we 640 // allocate will be freed at the end of the routine. 641 642 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 643 // but makes the entry a little slower. There is however a little dance we have to 644 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 645 ResetNoHandleMark rnhm; // No-op in release/product versions 646 HandleMark hm; 647 648 frame stub_frame = thread->last_frame(); 649 650 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 651 // must point to the vframeArray for the unpack frame. 652 vframeArray* array = thread->vframe_array_head(); 653 654 #ifndef PRODUCT 655 if (TraceDeoptimization) { 656 ttyLocker ttyl; 657 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", 658 p2i(thread), p2i(array), exec_mode); 659 } 660 #endif 661 Events::log_deopt_message(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 662 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode); 663 664 UnrollBlock* info = array->unroll_block(); 665 666 // We set the last_Java frame. But the stack isn't really parsable here. So we 667 // clear it to make sure JFR understands not to try and walk stacks from events 668 // in here. 669 intptr_t* sp = thread->frame_anchor()->last_Java_sp(); 670 thread->frame_anchor()->set_last_Java_sp(NULL); 671 672 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 673 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 674 675 thread->frame_anchor()->set_last_Java_sp(sp); 676 677 BasicType bt = info->return_type(); 678 679 // If we have an exception pending, claim that the return type is an oop 680 // so the deopt_blob does not overwrite the exception_oop. 681 682 if (exec_mode == Unpack_exception) 683 bt = T_OBJECT; 684 685 // Cleanup thread deopt data 686 cleanup_deopt_info(thread, array); 687 688 #ifndef PRODUCT 689 if (VerifyStack) { 690 ResourceMark res_mark; 691 // Clear pending exception to not break verification code (restored afterwards) 692 PRESERVE_EXCEPTION_MARK; 693 694 thread->validate_frame_layout(); 695 696 // Verify that the just-unpacked frames match the interpreter's 697 // notions of expression stack and locals 698 vframeArray* cur_array = thread->vframe_array_last(); 699 RegisterMap rm(thread, false); 700 rm.set_include_argument_oops(false); 701 bool is_top_frame = true; 702 int callee_size_of_parameters = 0; 703 int callee_max_locals = 0; 704 for (int i = 0; i < cur_array->frames(); i++) { 705 vframeArrayElement* el = cur_array->element(i); 706 frame* iframe = el->iframe(); 707 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 708 709 // Get the oop map for this bci 710 InterpreterOopMap mask; 711 int cur_invoke_parameter_size = 0; 712 bool try_next_mask = false; 713 int next_mask_expression_stack_size = -1; 714 int top_frame_expression_stack_adjustment = 0; 715 methodHandle mh(thread, iframe->interpreter_frame_method()); 716 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 717 BytecodeStream str(mh, iframe->interpreter_frame_bci()); 718 int max_bci = mh->code_size(); 719 // Get to the next bytecode if possible 720 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 721 // Check to see if we can grab the number of outgoing arguments 722 // at an uncommon trap for an invoke (where the compiler 723 // generates debug info before the invoke has executed) 724 Bytecodes::Code cur_code = str.next(); 725 if (Bytecodes::is_invoke(cur_code)) { 726 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 727 cur_invoke_parameter_size = invoke.size_of_parameters(); 728 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 729 callee_size_of_parameters++; 730 } 731 } 732 if (str.bci() < max_bci) { 733 Bytecodes::Code next_code = str.next(); 734 if (next_code >= 0) { 735 // The interpreter oop map generator reports results before 736 // the current bytecode has executed except in the case of 737 // calls. It seems to be hard to tell whether the compiler 738 // has emitted debug information matching the "state before" 739 // a given bytecode or the state after, so we try both 740 if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) { 741 // Get expression stack size for the next bytecode 742 InterpreterOopMap next_mask; 743 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 744 next_mask_expression_stack_size = next_mask.expression_stack_size(); 745 if (Bytecodes::is_invoke(next_code)) { 746 Bytecode_invoke invoke(mh, str.bci()); 747 next_mask_expression_stack_size += invoke.size_of_parameters(); 748 } 749 // Need to subtract off the size of the result type of 750 // the bytecode because this is not described in the 751 // debug info but returned to the interpreter in the TOS 752 // caching register 753 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 754 if (bytecode_result_type != T_ILLEGAL) { 755 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 756 } 757 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive"); 758 try_next_mask = true; 759 } 760 } 761 } 762 763 // Verify stack depth and oops in frame 764 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 765 if (!( 766 /* SPARC */ 767 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 768 /* x86 */ 769 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 770 (try_next_mask && 771 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 772 top_frame_expression_stack_adjustment))) || 773 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 774 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 775 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 776 )) { 777 { 778 ttyLocker ttyl; 779 780 // Print out some information that will help us debug the problem 781 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 782 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 783 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 784 iframe->interpreter_frame_expression_stack_size()); 785 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 786 tty->print_cr(" try_next_mask = %d", try_next_mask); 787 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 788 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 789 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 790 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 791 tty->print_cr(" exec_mode = %d", exec_mode); 792 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 793 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 794 tty->print_cr(" Interpreted frames:"); 795 for (int k = 0; k < cur_array->frames(); k++) { 796 vframeArrayElement* el = cur_array->element(k); 797 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 798 } 799 cur_array->print_on_2(tty); 800 } // release tty lock before calling guarantee 801 guarantee(false, "wrong number of expression stack elements during deopt"); 802 } 803 VerifyOopClosure verify; 804 iframe->oops_interpreted_do(&verify, &rm, false); 805 callee_size_of_parameters = mh->size_of_parameters(); 806 callee_max_locals = mh->max_locals(); 807 is_top_frame = false; 808 } 809 } 810 #endif /* !PRODUCT */ 811 812 return bt; 813 JRT_END 814 815 class DeoptimizeMarkedClosure : public HandshakeClosure { 816 public: 817 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {} 818 void do_thread(Thread* thread) { 819 JavaThread* jt = (JavaThread*)thread; 820 jt->deoptimize_marked_methods(); 821 } 822 }; 823 824 void Deoptimization::deoptimize_all_marked(nmethod* nmethod_only) { 825 ResourceMark rm; 826 DeoptimizationMarker dm; 827 828 // Make the dependent methods not entrant 829 if (nmethod_only != NULL) { 830 nmethod_only->mark_for_deoptimization(); 831 nmethod_only->make_not_entrant(); 832 } else { 833 MutexLocker mu(SafepointSynchronize::is_at_safepoint() ? NULL : CodeCache_lock, Mutex::_no_safepoint_check_flag); 834 CodeCache::make_marked_nmethods_not_entrant(); 835 } 836 837 DeoptimizeMarkedClosure deopt; 838 if (SafepointSynchronize::is_at_safepoint()) { 839 Threads::java_threads_do(&deopt); 840 } else { 841 Handshake::execute(&deopt); 842 } 843 } 844 845 Deoptimization::DeoptAction Deoptimization::_unloaded_action 846 = Deoptimization::Action_reinterpret; 847 848 849 850 #if INCLUDE_JVMCI || INCLUDE_AOT 851 template<typename CacheType> 852 class BoxCacheBase : public CHeapObj<mtCompiler> { 853 protected: 854 static InstanceKlass* find_cache_klass(Symbol* klass_name, TRAPS) { 855 ResourceMark rm; 856 char* klass_name_str = klass_name->as_C_string(); 857 Klass* k = SystemDictionary::find(klass_name, Handle(), Handle(), THREAD); 858 guarantee(k != NULL, "%s must be loaded", klass_name_str); 859 InstanceKlass* ik = InstanceKlass::cast(k); 860 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str); 861 CacheType::compute_offsets(ik); 862 return ik; 863 } 864 }; 865 866 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> { 867 PrimitiveType _low; 868 PrimitiveType _high; 869 jobject _cache; 870 protected: 871 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton; 872 BoxCache(Thread* thread) { 873 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(CacheType::symbol(), thread); 874 objArrayOop cache = CacheType::cache(ik); 875 assert(cache->length() > 0, "Empty cache"); 876 _low = BoxType::value(cache->obj_at(0)); 877 _high = _low + cache->length() - 1; 878 _cache = JNIHandles::make_global(Handle(thread, cache)); 879 } 880 ~BoxCache() { 881 JNIHandles::destroy_global(_cache); 882 } 883 public: 884 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) { 885 if (_singleton == NULL) { 886 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread); 887 if (!Atomic::replace_if_null(&_singleton, s)) { 888 delete s; 889 } 890 } 891 return _singleton; 892 } 893 oop lookup(PrimitiveType value) { 894 if (_low <= value && value <= _high) { 895 int offset = value - _low; 896 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset); 897 } 898 return NULL; 899 } 900 oop lookup_raw(intptr_t raw_value) { 901 // Have to cast to avoid little/big-endian problems. 902 if (sizeof(PrimitiveType) > sizeof(jint)) { 903 jlong value = (jlong)raw_value; 904 return lookup(value); 905 } 906 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value); 907 return lookup(value); 908 } 909 }; 910 911 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache; 912 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache; 913 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache; 914 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache; 915 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache; 916 917 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = NULL; 918 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = NULL; 919 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = NULL; 920 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = NULL; 921 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = NULL; 922 923 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> { 924 jobject _true_cache; 925 jobject _false_cache; 926 protected: 927 static BooleanBoxCache *_singleton; 928 BooleanBoxCache(Thread *thread) { 929 InstanceKlass* ik = find_cache_klass(java_lang_Boolean::symbol(), thread); 930 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik))); 931 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik))); 932 } 933 ~BooleanBoxCache() { 934 JNIHandles::destroy_global(_true_cache); 935 JNIHandles::destroy_global(_false_cache); 936 } 937 public: 938 static BooleanBoxCache* singleton(Thread* thread) { 939 if (_singleton == NULL) { 940 BooleanBoxCache* s = new BooleanBoxCache(thread); 941 if (!Atomic::replace_if_null(&_singleton, s)) { 942 delete s; 943 } 944 } 945 return _singleton; 946 } 947 oop lookup_raw(intptr_t raw_value) { 948 // Have to cast to avoid little/big-endian problems. 949 jboolean value = (jboolean)*((jint*)&raw_value); 950 return lookup(value); 951 } 952 oop lookup(jboolean value) { 953 if (value != 0) { 954 return JNIHandles::resolve_non_null(_true_cache); 955 } 956 return JNIHandles::resolve_non_null(_false_cache); 957 } 958 }; 959 960 BooleanBoxCache* BooleanBoxCache::_singleton = NULL; 961 962 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, TRAPS) { 963 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()()); 964 BasicType box_type = SystemDictionary::box_klass_type(k); 965 if (box_type != T_OBJECT) { 966 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0)); 967 switch(box_type) { 968 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 969 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 970 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 971 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 972 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 973 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 974 default:; 975 } 976 } 977 return NULL; 978 } 979 #endif // INCLUDE_JVMCI || INCLUDE_AOT 980 981 #if COMPILER2_OR_JVMCI 982 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) { 983 Handle pending_exception(THREAD, thread->pending_exception()); 984 const char* exception_file = thread->exception_file(); 985 int exception_line = thread->exception_line(); 986 thread->clear_pending_exception(); 987 988 bool failures = false; 989 990 for (int i = 0; i < objects->length(); i++) { 991 assert(objects->at(i)->is_object(), "invalid debug information"); 992 ObjectValue* sv = (ObjectValue*) objects->at(i); 993 994 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 995 oop obj = NULL; 996 997 if (k->is_instance_klass()) { 998 #if INCLUDE_JVMCI || INCLUDE_AOT 999 CompiledMethod* cm = fr->cb()->as_compiled_method_or_null(); 1000 if (cm->is_compiled_by_jvmci() && sv->is_auto_box()) { 1001 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv; 1002 obj = get_cached_box(abv, fr, reg_map, THREAD); 1003 if (obj != NULL) { 1004 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it. 1005 abv->set_cached(true); 1006 } 1007 } 1008 #endif // INCLUDE_JVMCI || INCLUDE_AOT 1009 InstanceKlass* ik = InstanceKlass::cast(k); 1010 if (obj == NULL) { 1011 obj = ik->allocate_instance(THREAD); 1012 } 1013 } else if (k->is_typeArray_klass()) { 1014 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1015 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 1016 int len = sv->field_size() / type2size[ak->element_type()]; 1017 obj = ak->allocate(len, THREAD); 1018 } else if (k->is_objArray_klass()) { 1019 ObjArrayKlass* ak = ObjArrayKlass::cast(k); 1020 obj = ak->allocate(sv->field_size(), THREAD); 1021 } 1022 1023 if (obj == NULL) { 1024 failures = true; 1025 } 1026 1027 assert(sv->value().is_null(), "redundant reallocation"); 1028 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception"); 1029 CLEAR_PENDING_EXCEPTION; 1030 sv->set_value(obj); 1031 } 1032 1033 if (failures) { 1034 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 1035 } else if (pending_exception.not_null()) { 1036 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 1037 } 1038 1039 return failures; 1040 } 1041 1042 #if INCLUDE_JVMCI 1043 /** 1044 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints), 1045 * we need to somehow be able to recover the actual kind to be able to write the correct 1046 * amount of bytes. 1047 * For that purpose, this method assumes that, for an entry spanning n bytes at index i, 1048 * the entries at index n + 1 to n + i are 'markers'. 1049 * For example, if we were writing a short at index 4 of a byte array of size 8, the 1050 * expected form of the array would be: 1051 * 1052 * {b0, b1, b2, b3, INT, marker, b6, b7} 1053 * 1054 * Thus, in order to get back the size of the entry, we simply need to count the number 1055 * of marked entries 1056 * 1057 * @param virtualArray the virtualized byte array 1058 * @param i index of the virtual entry we are recovering 1059 * @return The number of bytes the entry spans 1060 */ 1061 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) { 1062 int index = i; 1063 while (++index < virtualArray->field_size() && 1064 virtualArray->field_at(index)->is_marker()) {} 1065 return index - i; 1066 } 1067 1068 /** 1069 * If there was a guarantee for byte array to always start aligned to a long, we could 1070 * do a simple check on the parity of the index. Unfortunately, that is not always the 1071 * case. Thus, we check alignment of the actual address we are writing to. 1072 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to 1073 * write a long to index 3. 1074 */ 1075 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) { 1076 jbyte* res = obj->byte_at_addr(index); 1077 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write"); 1078 return res; 1079 } 1080 1081 static void byte_array_put(typeArrayOop obj, intptr_t val, int index, int byte_count) { 1082 switch (byte_count) { 1083 case 1: 1084 obj->byte_at_put(index, (jbyte) *((jint *) &val)); 1085 break; 1086 case 2: 1087 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) *((jint *) &val); 1088 break; 1089 case 4: 1090 *((jint *) check_alignment_get_addr(obj, index, 4)) = (jint) *((jint *) &val); 1091 break; 1092 case 8: 1093 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) *((jlong *) &val); 1094 break; 1095 default: 1096 ShouldNotReachHere(); 1097 } 1098 } 1099 #endif // INCLUDE_JVMCI 1100 1101 1102 // restore elements of an eliminated type array 1103 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 1104 int index = 0; 1105 intptr_t val; 1106 1107 for (int i = 0; i < sv->field_size(); i++) { 1108 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1109 switch(type) { 1110 case T_LONG: case T_DOUBLE: { 1111 assert(value->type() == T_INT, "Agreement."); 1112 StackValue* low = 1113 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1114 #ifdef _LP64 1115 jlong res = (jlong)low->get_int(); 1116 #else 1117 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1118 #endif 1119 obj->long_at_put(index, res); 1120 break; 1121 } 1122 1123 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1124 case T_INT: case T_FLOAT: { // 4 bytes. 1125 assert(value->type() == T_INT, "Agreement."); 1126 bool big_value = false; 1127 if (i + 1 < sv->field_size() && type == T_INT) { 1128 if (sv->field_at(i)->is_location()) { 1129 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 1130 if (type == Location::dbl || type == Location::lng) { 1131 big_value = true; 1132 } 1133 } else if (sv->field_at(i)->is_constant_int()) { 1134 ScopeValue* next_scope_field = sv->field_at(i + 1); 1135 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1136 big_value = true; 1137 } 1138 } 1139 } 1140 1141 if (big_value) { 1142 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1143 #ifdef _LP64 1144 jlong res = (jlong)low->get_int(); 1145 #else 1146 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1147 #endif 1148 obj->int_at_put(index, (jint)*((jint*)&res)); 1149 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1)); 1150 } else { 1151 val = value->get_int(); 1152 obj->int_at_put(index, (jint)*((jint*)&val)); 1153 } 1154 break; 1155 } 1156 1157 case T_SHORT: 1158 assert(value->type() == T_INT, "Agreement."); 1159 val = value->get_int(); 1160 obj->short_at_put(index, (jshort)*((jint*)&val)); 1161 break; 1162 1163 case T_CHAR: 1164 assert(value->type() == T_INT, "Agreement."); 1165 val = value->get_int(); 1166 obj->char_at_put(index, (jchar)*((jint*)&val)); 1167 break; 1168 1169 case T_BYTE: { 1170 assert(value->type() == T_INT, "Agreement."); 1171 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'. 1172 val = value->get_int(); 1173 #if INCLUDE_JVMCI 1174 int byte_count = count_number_of_bytes_for_entry(sv, i); 1175 byte_array_put(obj, val, index, byte_count); 1176 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip. 1177 i += byte_count - 1; // Balance the loop counter. 1178 index += byte_count; 1179 // index has been updated so continue at top of loop 1180 continue; 1181 #else 1182 obj->byte_at_put(index, (jbyte)*((jint*)&val)); 1183 break; 1184 #endif // INCLUDE_JVMCI 1185 } 1186 1187 case T_BOOLEAN: { 1188 assert(value->type() == T_INT, "Agreement."); 1189 val = value->get_int(); 1190 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 1191 break; 1192 } 1193 1194 default: 1195 ShouldNotReachHere(); 1196 } 1197 index++; 1198 } 1199 } 1200 1201 // restore fields of an eliminated object array 1202 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 1203 for (int i = 0; i < sv->field_size(); i++) { 1204 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1205 assert(value->type() == T_OBJECT, "object element expected"); 1206 obj->obj_at_put(i, value->get_obj()()); 1207 } 1208 } 1209 1210 class ReassignedField { 1211 public: 1212 int _offset; 1213 BasicType _type; 1214 public: 1215 ReassignedField() { 1216 _offset = 0; 1217 _type = T_ILLEGAL; 1218 } 1219 }; 1220 1221 int compare(ReassignedField* left, ReassignedField* right) { 1222 return left->_offset - right->_offset; 1223 } 1224 1225 // Restore fields of an eliminated instance object using the same field order 1226 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 1227 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) { 1228 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 1229 InstanceKlass* ik = klass; 1230 while (ik != NULL) { 1231 for (AllFieldStream fs(ik); !fs.done(); fs.next()) { 1232 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) { 1233 ReassignedField field; 1234 field._offset = fs.offset(); 1235 field._type = Signature::basic_type(fs.signature()); 1236 fields->append(field); 1237 } 1238 } 1239 ik = ik->superklass(); 1240 } 1241 fields->sort(compare); 1242 for (int i = 0; i < fields->length(); i++) { 1243 intptr_t val; 1244 ScopeValue* scope_field = sv->field_at(svIndex); 1245 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1246 int offset = fields->at(i)._offset; 1247 BasicType type = fields->at(i)._type; 1248 switch (type) { 1249 case T_OBJECT: case T_ARRAY: 1250 assert(value->type() == T_OBJECT, "Agreement."); 1251 obj->obj_field_put(offset, value->get_obj()()); 1252 break; 1253 1254 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1255 case T_INT: case T_FLOAT: { // 4 bytes. 1256 assert(value->type() == T_INT, "Agreement."); 1257 bool big_value = false; 1258 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1259 if (scope_field->is_location()) { 1260 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1261 if (type == Location::dbl || type == Location::lng) { 1262 big_value = true; 1263 } 1264 } 1265 if (scope_field->is_constant_int()) { 1266 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1267 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1268 big_value = true; 1269 } 1270 } 1271 } 1272 1273 if (big_value) { 1274 i++; 1275 assert(i < fields->length(), "second T_INT field needed"); 1276 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1277 } else { 1278 val = value->get_int(); 1279 obj->int_field_put(offset, (jint)*((jint*)&val)); 1280 break; 1281 } 1282 } 1283 /* no break */ 1284 1285 case T_LONG: case T_DOUBLE: { 1286 assert(value->type() == T_INT, "Agreement."); 1287 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1288 #ifdef _LP64 1289 jlong res = (jlong)low->get_int(); 1290 #else 1291 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1292 #endif 1293 obj->long_field_put(offset, res); 1294 break; 1295 } 1296 1297 case T_SHORT: 1298 assert(value->type() == T_INT, "Agreement."); 1299 val = value->get_int(); 1300 obj->short_field_put(offset, (jshort)*((jint*)&val)); 1301 break; 1302 1303 case T_CHAR: 1304 assert(value->type() == T_INT, "Agreement."); 1305 val = value->get_int(); 1306 obj->char_field_put(offset, (jchar)*((jint*)&val)); 1307 break; 1308 1309 case T_BYTE: 1310 assert(value->type() == T_INT, "Agreement."); 1311 val = value->get_int(); 1312 obj->byte_field_put(offset, (jbyte)*((jint*)&val)); 1313 break; 1314 1315 case T_BOOLEAN: 1316 assert(value->type() == T_INT, "Agreement."); 1317 val = value->get_int(); 1318 obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 1319 break; 1320 1321 default: 1322 ShouldNotReachHere(); 1323 } 1324 svIndex++; 1325 } 1326 return svIndex; 1327 } 1328 1329 // restore fields of all eliminated objects and arrays 1330 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) { 1331 for (int i = 0; i < objects->length(); i++) { 1332 ObjectValue* sv = (ObjectValue*) objects->at(i); 1333 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1334 Handle obj = sv->value(); 1335 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1336 if (PrintDeoptimizationDetails) { 1337 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1338 } 1339 if (obj.is_null()) { 1340 continue; 1341 } 1342 #if INCLUDE_JVMCI || INCLUDE_AOT 1343 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS. 1344 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) { 1345 continue; 1346 } 1347 #endif // INCLUDE_JVMCI || INCLUDE_AOT 1348 if (k->is_instance_klass()) { 1349 InstanceKlass* ik = InstanceKlass::cast(k); 1350 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal); 1351 } else if (k->is_typeArray_klass()) { 1352 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1353 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1354 } else if (k->is_objArray_klass()) { 1355 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1356 } 1357 } 1358 } 1359 1360 1361 // relock objects for which synchronization was eliminated 1362 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) { 1363 for (int i = 0; i < monitors->length(); i++) { 1364 MonitorInfo* mon_info = monitors->at(i); 1365 if (mon_info->eliminated()) { 1366 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1367 if (!mon_info->owner_is_scalar_replaced()) { 1368 Handle obj(thread, mon_info->owner()); 1369 markWord mark = obj->mark(); 1370 if (UseBiasedLocking && mark.has_bias_pattern()) { 1371 // New allocated objects may have the mark set to anonymously biased. 1372 // Also the deoptimized method may called methods with synchronization 1373 // where the thread-local object is bias locked to the current thread. 1374 assert(mark.is_biased_anonymously() || 1375 mark.biased_locker() == thread, "should be locked to current thread"); 1376 // Reset mark word to unbiased prototype. 1377 markWord unbiased_prototype = markWord::prototype().set_age(mark.age()); 1378 obj->set_mark(unbiased_prototype); 1379 } 1380 BasicLock* lock = mon_info->lock(); 1381 ObjectSynchronizer::enter(obj, lock, thread); 1382 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1383 } 1384 } 1385 } 1386 } 1387 1388 1389 #ifndef PRODUCT 1390 // print information about reallocated objects 1391 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 1392 fieldDescriptor fd; 1393 1394 for (int i = 0; i < objects->length(); i++) { 1395 ObjectValue* sv = (ObjectValue*) objects->at(i); 1396 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1397 Handle obj = sv->value(); 1398 1399 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 1400 k->print_value(); 1401 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1402 if (obj.is_null()) { 1403 tty->print(" allocation failed"); 1404 } else { 1405 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 1406 } 1407 tty->cr(); 1408 1409 if (Verbose && !obj.is_null()) { 1410 k->oop_print_on(obj(), tty); 1411 } 1412 } 1413 } 1414 #endif 1415 #endif // COMPILER2_OR_JVMCI 1416 1417 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1418 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1419 1420 #ifndef PRODUCT 1421 if (PrintDeoptimizationDetails) { 1422 ttyLocker ttyl; 1423 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread)); 1424 fr.print_on(tty); 1425 tty->print_cr(" Virtual frames (innermost first):"); 1426 for (int index = 0; index < chunk->length(); index++) { 1427 compiledVFrame* vf = chunk->at(index); 1428 tty->print(" %2d - ", index); 1429 vf->print_value(); 1430 int bci = chunk->at(index)->raw_bci(); 1431 const char* code_name; 1432 if (bci == SynchronizationEntryBCI) { 1433 code_name = "sync entry"; 1434 } else { 1435 Bytecodes::Code code = vf->method()->code_at(bci); 1436 code_name = Bytecodes::name(code); 1437 } 1438 tty->print(" - %s", code_name); 1439 tty->print_cr(" @ bci %d ", bci); 1440 if (Verbose) { 1441 vf->print(); 1442 tty->cr(); 1443 } 1444 } 1445 } 1446 #endif 1447 1448 // Register map for next frame (used for stack crawl). We capture 1449 // the state of the deopt'ing frame's caller. Thus if we need to 1450 // stuff a C2I adapter we can properly fill in the callee-save 1451 // register locations. 1452 frame caller = fr.sender(reg_map); 1453 int frame_size = caller.sp() - fr.sp(); 1454 1455 frame sender = caller; 1456 1457 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1458 // the vframeArray containing the unpacking information is allocated in the C heap. 1459 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1460 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1461 1462 // Compare the vframeArray to the collected vframes 1463 assert(array->structural_compare(thread, chunk), "just checking"); 1464 1465 #ifndef PRODUCT 1466 if (PrintDeoptimizationDetails) { 1467 ttyLocker ttyl; 1468 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array)); 1469 } 1470 #endif // PRODUCT 1471 1472 return array; 1473 } 1474 1475 #if COMPILER2_OR_JVMCI 1476 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1477 // Reallocation of some scalar replaced objects failed. Record 1478 // that we need to pop all the interpreter frames for the 1479 // deoptimized compiled frame. 1480 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1481 thread->set_frames_to_pop_failed_realloc(array->frames()); 1482 // Unlock all monitors here otherwise the interpreter will see a 1483 // mix of locked and unlocked monitors (because of failed 1484 // reallocations of synchronized objects) and be confused. 1485 for (int i = 0; i < array->frames(); i++) { 1486 MonitorChunk* monitors = array->element(i)->monitors(); 1487 if (monitors != NULL) { 1488 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1489 BasicObjectLock* src = monitors->at(j); 1490 if (src->obj() != NULL) { 1491 ObjectSynchronizer::exit(src->obj(), src->lock(), thread); 1492 } 1493 } 1494 array->element(i)->free_monitors(thread); 1495 #ifdef ASSERT 1496 array->element(i)->set_removed_monitors(); 1497 #endif 1498 } 1499 } 1500 } 1501 #endif 1502 1503 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1504 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1505 Thread* thread = Thread::current(); 1506 for (int i = 0; i < monitors->length(); i++) { 1507 MonitorInfo* mon_info = monitors->at(i); 1508 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1509 objects_to_revoke->append(Handle(thread, mon_info->owner())); 1510 } 1511 } 1512 } 1513 1514 static void get_monitors_from_stack(GrowableArray<Handle>* objects_to_revoke, JavaThread* thread, frame fr, RegisterMap* map) { 1515 // Unfortunately we don't have a RegisterMap available in most of 1516 // the places we want to call this routine so we need to walk the 1517 // stack again to update the register map. 1518 if (map == NULL || !map->update_map()) { 1519 StackFrameStream sfs(thread, true); 1520 bool found = false; 1521 while (!found && !sfs.is_done()) { 1522 frame* cur = sfs.current(); 1523 sfs.next(); 1524 found = cur->id() == fr.id(); 1525 } 1526 assert(found, "frame to be deoptimized not found on target thread's stack"); 1527 map = sfs.register_map(); 1528 } 1529 1530 vframe* vf = vframe::new_vframe(&fr, map, thread); 1531 compiledVFrame* cvf = compiledVFrame::cast(vf); 1532 // Revoke monitors' biases in all scopes 1533 while (!cvf->is_top()) { 1534 collect_monitors(cvf, objects_to_revoke); 1535 cvf = compiledVFrame::cast(cvf->sender()); 1536 } 1537 collect_monitors(cvf, objects_to_revoke); 1538 } 1539 1540 void Deoptimization::revoke_from_deopt_handler(JavaThread* thread, frame fr, RegisterMap* map) { 1541 if (!UseBiasedLocking) { 1542 return; 1543 } 1544 ResourceMark rm; 1545 HandleMark hm; 1546 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1547 get_monitors_from_stack(objects_to_revoke, thread, fr, map); 1548 1549 int len = objects_to_revoke->length(); 1550 for (int i = 0; i < len; i++) { 1551 oop obj = (objects_to_revoke->at(i))(); 1552 BiasedLocking::revoke_own_lock(objects_to_revoke->at(i), thread); 1553 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 1554 } 1555 } 1556 1557 1558 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1559 assert(fr.can_be_deoptimized(), "checking frame type"); 1560 1561 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1562 1563 if (LogCompilation && xtty != NULL) { 1564 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null(); 1565 assert(cm != NULL, "only compiled methods can deopt"); 1566 1567 ttyLocker ttyl; 1568 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1569 cm->log_identity(xtty); 1570 xtty->end_head(); 1571 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1572 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1573 xtty->method(sd->method()); 1574 xtty->end_elem(); 1575 if (sd->is_top()) break; 1576 } 1577 xtty->tail("deoptimized"); 1578 } 1579 1580 // Patch the compiled method so that when execution returns to it we will 1581 // deopt the execution state and return to the interpreter. 1582 fr.deoptimize(thread); 1583 } 1584 1585 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) { 1586 // Deoptimize only if the frame comes from compile code. 1587 // Do not deoptimize the frame which is already patched 1588 // during the execution of the loops below. 1589 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1590 return; 1591 } 1592 ResourceMark rm; 1593 DeoptimizationMarker dm; 1594 deoptimize_single_frame(thread, fr, reason); 1595 } 1596 1597 #if INCLUDE_JVMCI 1598 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) { 1599 // there is no exception handler for this pc => deoptimize 1600 cm->make_not_entrant(); 1601 1602 // Use Deoptimization::deoptimize for all of its side-effects: 1603 // gathering traps statistics, logging... 1604 // it also patches the return pc but we do not care about that 1605 // since we return a continuation to the deopt_blob below. 1606 JavaThread* thread = JavaThread::current(); 1607 RegisterMap reg_map(thread, false); 1608 frame runtime_frame = thread->last_frame(); 1609 frame caller_frame = runtime_frame.sender(®_map); 1610 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method"); 1611 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler); 1612 1613 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, cm->method()), true); 1614 if (trap_mdo != NULL) { 1615 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); 1616 } 1617 1618 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); 1619 } 1620 #endif 1621 1622 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1623 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1624 "can only deoptimize other thread at a safepoint"); 1625 // Compute frame and register map based on thread and sp. 1626 RegisterMap reg_map(thread, false); 1627 frame fr = thread->last_frame(); 1628 while (fr.id() != id) { 1629 fr = fr.sender(®_map); 1630 } 1631 deoptimize(thread, fr, reason); 1632 } 1633 1634 1635 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1636 if (thread == Thread::current()) { 1637 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1638 } else { 1639 VM_DeoptimizeFrame deopt(thread, id, reason); 1640 VMThread::execute(&deopt); 1641 } 1642 } 1643 1644 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1645 deoptimize_frame(thread, id, Reason_constraint); 1646 } 1647 1648 // JVMTI PopFrame support 1649 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1650 { 1651 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1652 } 1653 JRT_END 1654 1655 MethodData* 1656 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1657 bool create_if_missing) { 1658 Thread* THREAD = thread; 1659 MethodData* mdo = m()->method_data(); 1660 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1661 // Build an MDO. Ignore errors like OutOfMemory; 1662 // that simply means we won't have an MDO to update. 1663 Method::build_interpreter_method_data(m, THREAD); 1664 if (HAS_PENDING_EXCEPTION) { 1665 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1666 CLEAR_PENDING_EXCEPTION; 1667 } 1668 mdo = m()->method_data(); 1669 } 1670 return mdo; 1671 } 1672 1673 #if COMPILER2_OR_JVMCI 1674 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1675 // In case of an unresolved klass entry, load the class. 1676 // This path is exercised from case _ldc in Parse::do_one_bytecode, 1677 // and probably nowhere else. 1678 // Even that case would benefit from simply re-interpreting the 1679 // bytecode, without paying special attention to the class index. 1680 // So this whole "class index" feature should probably be removed. 1681 1682 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1683 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK); 1684 return; 1685 } 1686 1687 assert(!constant_pool->tag_at(index).is_symbol(), 1688 "no symbolic names here, please"); 1689 } 1690 1691 1692 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) { 1693 EXCEPTION_MARK; 1694 load_class_by_index(constant_pool, index, THREAD); 1695 if (HAS_PENDING_EXCEPTION) { 1696 // Exception happened during classloading. We ignore the exception here, since it 1697 // is going to be rethrown since the current activation is going to be deoptimized and 1698 // the interpreter will re-execute the bytecode. 1699 CLEAR_PENDING_EXCEPTION; 1700 // Class loading called java code which may have caused a stack 1701 // overflow. If the exception was thrown right before the return 1702 // to the runtime the stack is no longer guarded. Reguard the 1703 // stack otherwise if we return to the uncommon trap blob and the 1704 // stack bang causes a stack overflow we crash. 1705 assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1706 JavaThread* thread = (JavaThread*)THREAD; 1707 bool guard_pages_enabled = thread->stack_guards_enabled(); 1708 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1709 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1710 } 1711 } 1712 1713 #if INCLUDE_JFR 1714 1715 class DeoptReasonSerializer : public JfrSerializer { 1716 public: 1717 void serialize(JfrCheckpointWriter& writer) { 1718 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1) 1719 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) { 1720 writer.write_key((u8)i); 1721 writer.write(Deoptimization::trap_reason_name(i)); 1722 } 1723 } 1724 }; 1725 1726 class DeoptActionSerializer : public JfrSerializer { 1727 public: 1728 void serialize(JfrCheckpointWriter& writer) { 1729 static const u4 nof_actions = Deoptimization::Action_LIMIT; 1730 writer.write_count(nof_actions); 1731 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) { 1732 writer.write_key(i); 1733 writer.write(Deoptimization::trap_action_name((int)i)); 1734 } 1735 } 1736 }; 1737 1738 static void register_serializers() { 1739 static int critical_section = 0; 1740 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) { 1741 return; 1742 } 1743 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer()); 1744 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer()); 1745 } 1746 1747 static void post_deoptimization_event(CompiledMethod* nm, 1748 const Method* method, 1749 int trap_bci, 1750 int instruction, 1751 Deoptimization::DeoptReason reason, 1752 Deoptimization::DeoptAction action) { 1753 assert(nm != NULL, "invariant"); 1754 assert(method != NULL, "invariant"); 1755 if (EventDeoptimization::is_enabled()) { 1756 static bool serializers_registered = false; 1757 if (!serializers_registered) { 1758 register_serializers(); 1759 serializers_registered = true; 1760 } 1761 EventDeoptimization event; 1762 event.set_compileId(nm->compile_id()); 1763 event.set_compiler(nm->compiler_type()); 1764 event.set_method(method); 1765 event.set_lineNumber(method->line_number_from_bci(trap_bci)); 1766 event.set_bci(trap_bci); 1767 event.set_instruction(instruction); 1768 event.set_reason(reason); 1769 event.set_action(action); 1770 event.commit(); 1771 } 1772 } 1773 1774 #endif // INCLUDE_JFR 1775 1776 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1777 HandleMark hm; 1778 1779 // uncommon_trap() is called at the beginning of the uncommon trap 1780 // handler. Note this fact before we start generating temporary frames 1781 // that can confuse an asynchronous stack walker. This counter is 1782 // decremented at the end of unpack_frames(). 1783 thread->inc_in_deopt_handler(); 1784 1785 // We need to update the map if we have biased locking. 1786 #if INCLUDE_JVMCI 1787 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 1788 RegisterMap reg_map(thread, true); 1789 #else 1790 RegisterMap reg_map(thread, UseBiasedLocking); 1791 #endif 1792 frame stub_frame = thread->last_frame(); 1793 frame fr = stub_frame.sender(®_map); 1794 // Make sure the calling nmethod is not getting deoptimized and removed 1795 // before we are done with it. 1796 nmethodLocker nl(fr.pc()); 1797 1798 // Log a message 1799 Events::log_deopt_message(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 1800 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 1801 1802 { 1803 ResourceMark rm; 1804 1805 DeoptReason reason = trap_request_reason(trap_request); 1806 DeoptAction action = trap_request_action(trap_request); 1807 #if INCLUDE_JVMCI 1808 int debug_id = trap_request_debug_id(trap_request); 1809 #endif 1810 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1811 1812 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1813 compiledVFrame* cvf = compiledVFrame::cast(vf); 1814 1815 CompiledMethod* nm = cvf->code(); 1816 1817 ScopeDesc* trap_scope = cvf->scope(); 1818 1819 if (TraceDeoptimization) { 1820 ttyLocker ttyl; 1821 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"), trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string() 1822 #if INCLUDE_JVMCI 1823 , debug_id 1824 #endif 1825 ); 1826 } 1827 1828 methodHandle trap_method(THREAD, trap_scope->method()); 1829 int trap_bci = trap_scope->bci(); 1830 #if INCLUDE_JVMCI 1831 jlong speculation = thread->pending_failed_speculation(); 1832 if (nm->is_compiled_by_jvmci() && nm->is_nmethod()) { // Exclude AOTed methods 1833 nm->as_nmethod()->update_speculation(thread); 1834 } else { 1835 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers"); 1836 } 1837 1838 if (trap_bci == SynchronizationEntryBCI) { 1839 trap_bci = 0; 1840 thread->set_pending_monitorenter(true); 1841 } 1842 1843 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 1844 thread->set_pending_transfer_to_interpreter(true); 1845 } 1846 #endif 1847 1848 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1849 // Record this event in the histogram. 1850 gather_statistics(reason, action, trap_bc); 1851 1852 // Ensure that we can record deopt. history: 1853 // Need MDO to record RTM code generation state. 1854 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking ); 1855 1856 methodHandle profiled_method; 1857 #if INCLUDE_JVMCI 1858 if (nm->is_compiled_by_jvmci()) { 1859 profiled_method = methodHandle(THREAD, nm->method()); 1860 } else { 1861 profiled_method = trap_method; 1862 } 1863 #else 1864 profiled_method = trap_method; 1865 #endif 1866 1867 MethodData* trap_mdo = 1868 get_method_data(thread, profiled_method, create_if_missing); 1869 1870 JFR_ONLY(post_deoptimization_event(nm, trap_method(), trap_bci, trap_bc, reason, action);) 1871 1872 // Log a message 1873 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s", 1874 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()), 1875 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); 1876 1877 // Print a bunch of diagnostics, if requested. 1878 if (TraceDeoptimization || LogCompilation) { 1879 ResourceMark rm; 1880 ttyLocker ttyl; 1881 char buf[100]; 1882 if (xtty != NULL) { 1883 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 1884 os::current_thread_id(), 1885 format_trap_request(buf, sizeof(buf), trap_request)); 1886 #if INCLUDE_JVMCI 1887 if (speculation != 0) { 1888 xtty->print(" speculation='" JLONG_FORMAT "'", speculation); 1889 } 1890 #endif 1891 nm->log_identity(xtty); 1892 } 1893 Symbol* class_name = NULL; 1894 bool unresolved = false; 1895 if (unloaded_class_index >= 0) { 1896 constantPoolHandle constants (THREAD, trap_method->constants()); 1897 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1898 class_name = constants->klass_name_at(unloaded_class_index); 1899 unresolved = true; 1900 if (xtty != NULL) 1901 xtty->print(" unresolved='1'"); 1902 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1903 class_name = constants->symbol_at(unloaded_class_index); 1904 } 1905 if (xtty != NULL) 1906 xtty->name(class_name); 1907 } 1908 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) { 1909 // Dump the relevant MDO state. 1910 // This is the deopt count for the current reason, any previous 1911 // reasons or recompiles seen at this point. 1912 int dcnt = trap_mdo->trap_count(reason); 1913 if (dcnt != 0) 1914 xtty->print(" count='%d'", dcnt); 1915 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1916 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1917 if (dos != 0) { 1918 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1919 if (trap_state_is_recompiled(dos)) { 1920 int recnt2 = trap_mdo->overflow_recompile_count(); 1921 if (recnt2 != 0) 1922 xtty->print(" recompiles2='%d'", recnt2); 1923 } 1924 } 1925 } 1926 if (xtty != NULL) { 1927 xtty->stamp(); 1928 xtty->end_head(); 1929 } 1930 if (TraceDeoptimization) { // make noise on the tty 1931 tty->print("Uncommon trap occurred in"); 1932 nm->method()->print_short_name(tty); 1933 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id()); 1934 #if INCLUDE_JVMCI 1935 if (nm->is_nmethod()) { 1936 const char* installed_code_name = nm->as_nmethod()->jvmci_name(); 1937 if (installed_code_name != NULL) { 1938 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name); 1939 } 1940 } 1941 #endif 1942 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 1943 p2i(fr.pc()), 1944 os::current_thread_id(), 1945 trap_reason_name(reason), 1946 trap_action_name(action), 1947 unloaded_class_index 1948 #if INCLUDE_JVMCI 1949 , debug_id 1950 #endif 1951 ); 1952 if (class_name != NULL) { 1953 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1954 class_name->print_symbol_on(tty); 1955 } 1956 tty->cr(); 1957 } 1958 if (xtty != NULL) { 1959 // Log the precise location of the trap. 1960 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1961 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1962 xtty->method(sd->method()); 1963 xtty->end_elem(); 1964 if (sd->is_top()) break; 1965 } 1966 xtty->tail("uncommon_trap"); 1967 } 1968 } 1969 // (End diagnostic printout.) 1970 1971 // Load class if necessary 1972 if (unloaded_class_index >= 0) { 1973 constantPoolHandle constants(THREAD, trap_method->constants()); 1974 load_class_by_index(constants, unloaded_class_index); 1975 } 1976 1977 // Flush the nmethod if necessary and desirable. 1978 // 1979 // We need to avoid situations where we are re-flushing the nmethod 1980 // because of a hot deoptimization site. Repeated flushes at the same 1981 // point need to be detected by the compiler and avoided. If the compiler 1982 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1983 // module must take measures to avoid an infinite cycle of recompilation 1984 // and deoptimization. There are several such measures: 1985 // 1986 // 1. If a recompilation is ordered a second time at some site X 1987 // and for the same reason R, the action is adjusted to 'reinterpret', 1988 // to give the interpreter time to exercise the method more thoroughly. 1989 // If this happens, the method's overflow_recompile_count is incremented. 1990 // 1991 // 2. If the compiler fails to reduce the deoptimization rate, then 1992 // the method's overflow_recompile_count will begin to exceed the set 1993 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1994 // is adjusted to 'make_not_compilable', and the method is abandoned 1995 // to the interpreter. This is a performance hit for hot methods, 1996 // but is better than a disastrous infinite cycle of recompilations. 1997 // (Actually, only the method containing the site X is abandoned.) 1998 // 1999 // 3. In parallel with the previous measures, if the total number of 2000 // recompilations of a method exceeds the much larger set limit 2001 // PerMethodRecompilationCutoff, the method is abandoned. 2002 // This should only happen if the method is very large and has 2003 // many "lukewarm" deoptimizations. The code which enforces this 2004 // limit is elsewhere (class nmethod, class Method). 2005 // 2006 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 2007 // to recompile at each bytecode independently of the per-BCI cutoff. 2008 // 2009 // The decision to update code is up to the compiler, and is encoded 2010 // in the Action_xxx code. If the compiler requests Action_none 2011 // no trap state is changed, no compiled code is changed, and the 2012 // computation suffers along in the interpreter. 2013 // 2014 // The other action codes specify various tactics for decompilation 2015 // and recompilation. Action_maybe_recompile is the loosest, and 2016 // allows the compiled code to stay around until enough traps are seen, 2017 // and until the compiler gets around to recompiling the trapping method. 2018 // 2019 // The other actions cause immediate removal of the present code. 2020 2021 // Traps caused by injected profile shouldn't pollute trap counts. 2022 bool injected_profile_trap = trap_method->has_injected_profile() && 2023 (reason == Reason_intrinsic || reason == Reason_unreached); 2024 2025 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 2026 bool make_not_entrant = false; 2027 bool make_not_compilable = false; 2028 bool reprofile = false; 2029 switch (action) { 2030 case Action_none: 2031 // Keep the old code. 2032 update_trap_state = false; 2033 break; 2034 case Action_maybe_recompile: 2035 // Do not need to invalidate the present code, but we can 2036 // initiate another 2037 // Start compiler without (necessarily) invalidating the nmethod. 2038 // The system will tolerate the old code, but new code should be 2039 // generated when possible. 2040 break; 2041 case Action_reinterpret: 2042 // Go back into the interpreter for a while, and then consider 2043 // recompiling form scratch. 2044 make_not_entrant = true; 2045 // Reset invocation counter for outer most method. 2046 // This will allow the interpreter to exercise the bytecodes 2047 // for a while before recompiling. 2048 // By contrast, Action_make_not_entrant is immediate. 2049 // 2050 // Note that the compiler will track null_check, null_assert, 2051 // range_check, and class_check events and log them as if they 2052 // had been traps taken from compiled code. This will update 2053 // the MDO trap history so that the next compilation will 2054 // properly detect hot trap sites. 2055 reprofile = true; 2056 break; 2057 case Action_make_not_entrant: 2058 // Request immediate recompilation, and get rid of the old code. 2059 // Make them not entrant, so next time they are called they get 2060 // recompiled. Unloaded classes are loaded now so recompile before next 2061 // time they are called. Same for uninitialized. The interpreter will 2062 // link the missing class, if any. 2063 make_not_entrant = true; 2064 break; 2065 case Action_make_not_compilable: 2066 // Give up on compiling this method at all. 2067 make_not_entrant = true; 2068 make_not_compilable = true; 2069 break; 2070 default: 2071 ShouldNotReachHere(); 2072 } 2073 2074 // Setting +ProfileTraps fixes the following, on all platforms: 2075 // 4852688: ProfileInterpreter is off by default for ia64. The result is 2076 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 2077 // recompile relies on a MethodData* to record heroic opt failures. 2078 2079 // Whether the interpreter is producing MDO data or not, we also need 2080 // to use the MDO to detect hot deoptimization points and control 2081 // aggressive optimization. 2082 bool inc_recompile_count = false; 2083 ProfileData* pdata = NULL; 2084 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) { 2085 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity"); 2086 uint this_trap_count = 0; 2087 bool maybe_prior_trap = false; 2088 bool maybe_prior_recompile = false; 2089 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 2090 #if INCLUDE_JVMCI 2091 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 2092 #endif 2093 nm->method(), 2094 //outputs: 2095 this_trap_count, 2096 maybe_prior_trap, 2097 maybe_prior_recompile); 2098 // Because the interpreter also counts null, div0, range, and class 2099 // checks, these traps from compiled code are double-counted. 2100 // This is harmless; it just means that the PerXTrapLimit values 2101 // are in effect a little smaller than they look. 2102 2103 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2104 if (per_bc_reason != Reason_none) { 2105 // Now take action based on the partially known per-BCI history. 2106 if (maybe_prior_trap 2107 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 2108 // If there are too many traps at this BCI, force a recompile. 2109 // This will allow the compiler to see the limit overflow, and 2110 // take corrective action, if possible. The compiler generally 2111 // does not use the exact PerBytecodeTrapLimit value, but instead 2112 // changes its tactics if it sees any traps at all. This provides 2113 // a little hysteresis, delaying a recompile until a trap happens 2114 // several times. 2115 // 2116 // Actually, since there is only one bit of counter per BCI, 2117 // the possible per-BCI counts are {0,1,(per-method count)}. 2118 // This produces accurate results if in fact there is only 2119 // one hot trap site, but begins to get fuzzy if there are 2120 // many sites. For example, if there are ten sites each 2121 // trapping two or more times, they each get the blame for 2122 // all of their traps. 2123 make_not_entrant = true; 2124 } 2125 2126 // Detect repeated recompilation at the same BCI, and enforce a limit. 2127 if (make_not_entrant && maybe_prior_recompile) { 2128 // More than one recompile at this point. 2129 inc_recompile_count = maybe_prior_trap; 2130 } 2131 } else { 2132 // For reasons which are not recorded per-bytecode, we simply 2133 // force recompiles unconditionally. 2134 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 2135 make_not_entrant = true; 2136 } 2137 2138 // Go back to the compiler if there are too many traps in this method. 2139 if (this_trap_count >= per_method_trap_limit(reason)) { 2140 // If there are too many traps in this method, force a recompile. 2141 // This will allow the compiler to see the limit overflow, and 2142 // take corrective action, if possible. 2143 // (This condition is an unlikely backstop only, because the 2144 // PerBytecodeTrapLimit is more likely to take effect first, 2145 // if it is applicable.) 2146 make_not_entrant = true; 2147 } 2148 2149 // Here's more hysteresis: If there has been a recompile at 2150 // this trap point already, run the method in the interpreter 2151 // for a while to exercise it more thoroughly. 2152 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 2153 reprofile = true; 2154 } 2155 } 2156 2157 // Take requested actions on the method: 2158 2159 // Recompile 2160 if (make_not_entrant) { 2161 if (!nm->make_not_entrant()) { 2162 return; // the call did not change nmethod's state 2163 } 2164 2165 if (pdata != NULL) { 2166 // Record the recompilation event, if any. 2167 int tstate0 = pdata->trap_state(); 2168 int tstate1 = trap_state_set_recompiled(tstate0, true); 2169 if (tstate1 != tstate0) 2170 pdata->set_trap_state(tstate1); 2171 } 2172 2173 #if INCLUDE_RTM_OPT 2174 // Restart collecting RTM locking abort statistic if the method 2175 // is recompiled for a reason other than RTM state change. 2176 // Assume that in new recompiled code the statistic could be different, 2177 // for example, due to different inlining. 2178 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 2179 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) { 2180 trap_mdo->atomic_set_rtm_state(ProfileRTM); 2181 } 2182 #endif 2183 // For code aging we count traps separately here, using make_not_entrant() 2184 // as a guard against simultaneous deopts in multiple threads. 2185 if (reason == Reason_tenured && trap_mdo != NULL) { 2186 trap_mdo->inc_tenure_traps(); 2187 } 2188 } 2189 2190 if (inc_recompile_count) { 2191 trap_mdo->inc_overflow_recompile_count(); 2192 if ((uint)trap_mdo->overflow_recompile_count() > 2193 (uint)PerBytecodeRecompilationCutoff) { 2194 // Give up on the method containing the bad BCI. 2195 if (trap_method() == nm->method()) { 2196 make_not_compilable = true; 2197 } else { 2198 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization); 2199 // But give grace to the enclosing nm->method(). 2200 } 2201 } 2202 } 2203 2204 // Reprofile 2205 if (reprofile) { 2206 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 2207 } 2208 2209 // Give up compiling 2210 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 2211 assert(make_not_entrant, "consistent"); 2212 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization); 2213 } 2214 2215 } // Free marked resources 2216 2217 } 2218 JRT_END 2219 2220 ProfileData* 2221 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2222 int trap_bci, 2223 Deoptimization::DeoptReason reason, 2224 bool update_total_trap_count, 2225 #if INCLUDE_JVMCI 2226 bool is_osr, 2227 #endif 2228 Method* compiled_method, 2229 //outputs: 2230 uint& ret_this_trap_count, 2231 bool& ret_maybe_prior_trap, 2232 bool& ret_maybe_prior_recompile) { 2233 bool maybe_prior_trap = false; 2234 bool maybe_prior_recompile = false; 2235 uint this_trap_count = 0; 2236 if (update_total_trap_count) { 2237 uint idx = reason; 2238 #if INCLUDE_JVMCI 2239 if (is_osr) { 2240 idx += Reason_LIMIT; 2241 } 2242 #endif 2243 uint prior_trap_count = trap_mdo->trap_count(idx); 2244 this_trap_count = trap_mdo->inc_trap_count(idx); 2245 2246 // If the runtime cannot find a place to store trap history, 2247 // it is estimated based on the general condition of the method. 2248 // If the method has ever been recompiled, or has ever incurred 2249 // a trap with the present reason , then this BCI is assumed 2250 // (pessimistically) to be the culprit. 2251 maybe_prior_trap = (prior_trap_count != 0); 2252 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2253 } 2254 ProfileData* pdata = NULL; 2255 2256 2257 // For reasons which are recorded per bytecode, we check per-BCI data. 2258 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2259 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2260 if (per_bc_reason != Reason_none) { 2261 // Find the profile data for this BCI. If there isn't one, 2262 // try to allocate one from the MDO's set of spares. 2263 // This will let us detect a repeated trap at this point. 2264 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 2265 2266 if (pdata != NULL) { 2267 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2268 if (LogCompilation && xtty != NULL) { 2269 ttyLocker ttyl; 2270 // no more room for speculative traps in this MDO 2271 xtty->elem("speculative_traps_oom"); 2272 } 2273 } 2274 // Query the trap state of this profile datum. 2275 int tstate0 = pdata->trap_state(); 2276 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2277 maybe_prior_trap = false; 2278 if (!trap_state_is_recompiled(tstate0)) 2279 maybe_prior_recompile = false; 2280 2281 // Update the trap state of this profile datum. 2282 int tstate1 = tstate0; 2283 // Record the reason. 2284 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2285 // Store the updated state on the MDO, for next time. 2286 if (tstate1 != tstate0) 2287 pdata->set_trap_state(tstate1); 2288 } else { 2289 if (LogCompilation && xtty != NULL) { 2290 ttyLocker ttyl; 2291 // Missing MDP? Leave a small complaint in the log. 2292 xtty->elem("missing_mdp bci='%d'", trap_bci); 2293 } 2294 } 2295 } 2296 2297 // Return results: 2298 ret_this_trap_count = this_trap_count; 2299 ret_maybe_prior_trap = maybe_prior_trap; 2300 ret_maybe_prior_recompile = maybe_prior_recompile; 2301 return pdata; 2302 } 2303 2304 void 2305 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2306 ResourceMark rm; 2307 // Ignored outputs: 2308 uint ignore_this_trap_count; 2309 bool ignore_maybe_prior_trap; 2310 bool ignore_maybe_prior_recompile; 2311 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2312 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2313 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler); 2314 query_update_method_data(trap_mdo, trap_bci, 2315 (DeoptReason)reason, 2316 update_total_counts, 2317 #if INCLUDE_JVMCI 2318 false, 2319 #endif 2320 NULL, 2321 ignore_this_trap_count, 2322 ignore_maybe_prior_trap, 2323 ignore_maybe_prior_recompile); 2324 } 2325 2326 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) { 2327 if (TraceDeoptimization) { 2328 tty->print("Uncommon trap "); 2329 } 2330 // Still in Java no safepoints 2331 { 2332 // This enters VM and may safepoint 2333 uncommon_trap_inner(thread, trap_request); 2334 } 2335 return fetch_unroll_info_helper(thread, exec_mode); 2336 } 2337 2338 // Local derived constants. 2339 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2340 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1; 2341 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2342 2343 //---------------------------trap_state_reason--------------------------------- 2344 Deoptimization::DeoptReason 2345 Deoptimization::trap_state_reason(int trap_state) { 2346 // This assert provides the link between the width of DataLayout::trap_bits 2347 // and the encoding of "recorded" reasons. It ensures there are enough 2348 // bits to store all needed reasons in the per-BCI MDO profile. 2349 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2350 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2351 trap_state -= recompile_bit; 2352 if (trap_state == DS_REASON_MASK) { 2353 return Reason_many; 2354 } else { 2355 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2356 return (DeoptReason)trap_state; 2357 } 2358 } 2359 //-------------------------trap_state_has_reason------------------------------- 2360 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2361 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2362 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2363 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2364 trap_state -= recompile_bit; 2365 if (trap_state == DS_REASON_MASK) { 2366 return -1; // true, unspecifically (bottom of state lattice) 2367 } else if (trap_state == reason) { 2368 return 1; // true, definitely 2369 } else if (trap_state == 0) { 2370 return 0; // false, definitely (top of state lattice) 2371 } else { 2372 return 0; // false, definitely 2373 } 2374 } 2375 //-------------------------trap_state_add_reason------------------------------- 2376 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2377 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2378 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2379 trap_state -= recompile_bit; 2380 if (trap_state == DS_REASON_MASK) { 2381 return trap_state + recompile_bit; // already at state lattice bottom 2382 } else if (trap_state == reason) { 2383 return trap_state + recompile_bit; // the condition is already true 2384 } else if (trap_state == 0) { 2385 return reason + recompile_bit; // no condition has yet been true 2386 } else { 2387 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2388 } 2389 } 2390 //-----------------------trap_state_is_recompiled------------------------------ 2391 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2392 return (trap_state & DS_RECOMPILE_BIT) != 0; 2393 } 2394 //-----------------------trap_state_set_recompiled----------------------------- 2395 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2396 if (z) return trap_state | DS_RECOMPILE_BIT; 2397 else return trap_state & ~DS_RECOMPILE_BIT; 2398 } 2399 //---------------------------format_trap_state--------------------------------- 2400 // This is used for debugging and diagnostics, including LogFile output. 2401 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2402 int trap_state) { 2403 assert(buflen > 0, "sanity"); 2404 DeoptReason reason = trap_state_reason(trap_state); 2405 bool recomp_flag = trap_state_is_recompiled(trap_state); 2406 // Re-encode the state from its decoded components. 2407 int decoded_state = 0; 2408 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2409 decoded_state = trap_state_add_reason(decoded_state, reason); 2410 if (recomp_flag) 2411 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2412 // If the state re-encodes properly, format it symbolically. 2413 // Because this routine is used for debugging and diagnostics, 2414 // be robust even if the state is a strange value. 2415 size_t len; 2416 if (decoded_state != trap_state) { 2417 // Random buggy state that doesn't decode?? 2418 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2419 } else { 2420 len = jio_snprintf(buf, buflen, "%s%s", 2421 trap_reason_name(reason), 2422 recomp_flag ? " recompiled" : ""); 2423 } 2424 return buf; 2425 } 2426 2427 2428 //--------------------------------statics-------------------------------------- 2429 const char* Deoptimization::_trap_reason_name[] = { 2430 // Note: Keep this in sync. with enum DeoptReason. 2431 "none", 2432 "null_check", 2433 "null_assert" JVMCI_ONLY("_or_unreached0"), 2434 "range_check", 2435 "class_check", 2436 "array_check", 2437 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2438 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2439 "profile_predicate", 2440 "unloaded", 2441 "uninitialized", 2442 "initialized", 2443 "unreached", 2444 "unhandled", 2445 "constraint", 2446 "div0_check", 2447 "age", 2448 "predicate", 2449 "loop_limit_check", 2450 "speculate_class_check", 2451 "speculate_null_check", 2452 "speculate_null_assert", 2453 "rtm_state_change", 2454 "unstable_if", 2455 "unstable_fused_if", 2456 #if INCLUDE_JVMCI 2457 "aliasing", 2458 "transfer_to_interpreter", 2459 "not_compiled_exception_handler", 2460 "unresolved", 2461 "jsr_mismatch", 2462 #endif 2463 "tenured" 2464 }; 2465 const char* Deoptimization::_trap_action_name[] = { 2466 // Note: Keep this in sync. with enum DeoptAction. 2467 "none", 2468 "maybe_recompile", 2469 "reinterpret", 2470 "make_not_entrant", 2471 "make_not_compilable" 2472 }; 2473 2474 const char* Deoptimization::trap_reason_name(int reason) { 2475 // Check that every reason has a name 2476 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2477 2478 if (reason == Reason_many) return "many"; 2479 if ((uint)reason < Reason_LIMIT) 2480 return _trap_reason_name[reason]; 2481 static char buf[20]; 2482 sprintf(buf, "reason%d", reason); 2483 return buf; 2484 } 2485 const char* Deoptimization::trap_action_name(int action) { 2486 // Check that every action has a name 2487 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2488 2489 if ((uint)action < Action_LIMIT) 2490 return _trap_action_name[action]; 2491 static char buf[20]; 2492 sprintf(buf, "action%d", action); 2493 return buf; 2494 } 2495 2496 // This is used for debugging and diagnostics, including LogFile output. 2497 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2498 int trap_request) { 2499 jint unloaded_class_index = trap_request_index(trap_request); 2500 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2501 const char* action = trap_action_name(trap_request_action(trap_request)); 2502 #if INCLUDE_JVMCI 2503 int debug_id = trap_request_debug_id(trap_request); 2504 #endif 2505 size_t len; 2506 if (unloaded_class_index < 0) { 2507 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2508 reason, action 2509 #if INCLUDE_JVMCI 2510 ,debug_id 2511 #endif 2512 ); 2513 } else { 2514 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2515 reason, action, unloaded_class_index 2516 #if INCLUDE_JVMCI 2517 ,debug_id 2518 #endif 2519 ); 2520 } 2521 return buf; 2522 } 2523 2524 juint Deoptimization::_deoptimization_hist 2525 [Deoptimization::Reason_LIMIT] 2526 [1 + Deoptimization::Action_LIMIT] 2527 [Deoptimization::BC_CASE_LIMIT] 2528 = {0}; 2529 2530 enum { 2531 LSB_BITS = 8, 2532 LSB_MASK = right_n_bits(LSB_BITS) 2533 }; 2534 2535 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2536 Bytecodes::Code bc) { 2537 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2538 assert(action >= 0 && action < Action_LIMIT, "oob"); 2539 _deoptimization_hist[Reason_none][0][0] += 1; // total 2540 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2541 juint* cases = _deoptimization_hist[reason][1+action]; 2542 juint* bc_counter_addr = NULL; 2543 juint bc_counter = 0; 2544 // Look for an unused counter, or an exact match to this BC. 2545 if (bc != Bytecodes::_illegal) { 2546 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2547 juint* counter_addr = &cases[bc_case]; 2548 juint counter = *counter_addr; 2549 if ((counter == 0 && bc_counter_addr == NULL) 2550 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2551 // this counter is either free or is already devoted to this BC 2552 bc_counter_addr = counter_addr; 2553 bc_counter = counter | bc; 2554 } 2555 } 2556 } 2557 if (bc_counter_addr == NULL) { 2558 // Overflow, or no given bytecode. 2559 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2560 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2561 } 2562 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2563 } 2564 2565 jint Deoptimization::total_deoptimization_count() { 2566 return _deoptimization_hist[Reason_none][0][0]; 2567 } 2568 2569 void Deoptimization::print_statistics() { 2570 juint total = total_deoptimization_count(); 2571 juint account = total; 2572 if (total != 0) { 2573 ttyLocker ttyl; 2574 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 2575 tty->print_cr("Deoptimization traps recorded:"); 2576 #define PRINT_STAT_LINE(name, r) \ 2577 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2578 PRINT_STAT_LINE("total", total); 2579 // For each non-zero entry in the histogram, print the reason, 2580 // the action, and (if specifically known) the type of bytecode. 2581 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2582 for (int action = 0; action < Action_LIMIT; action++) { 2583 juint* cases = _deoptimization_hist[reason][1+action]; 2584 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2585 juint counter = cases[bc_case]; 2586 if (counter != 0) { 2587 char name[1*K]; 2588 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2589 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2590 bc = Bytecodes::_illegal; 2591 sprintf(name, "%s/%s/%s", 2592 trap_reason_name(reason), 2593 trap_action_name(action), 2594 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2595 juint r = counter >> LSB_BITS; 2596 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2597 account -= r; 2598 } 2599 } 2600 } 2601 } 2602 if (account != 0) { 2603 PRINT_STAT_LINE("unaccounted", account); 2604 } 2605 #undef PRINT_STAT_LINE 2606 if (xtty != NULL) xtty->tail("statistics"); 2607 } 2608 } 2609 #else // COMPILER2_OR_JVMCI 2610 2611 2612 // Stubs for C1 only system. 2613 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2614 return false; 2615 } 2616 2617 const char* Deoptimization::trap_reason_name(int reason) { 2618 return "unknown"; 2619 } 2620 2621 void Deoptimization::print_statistics() { 2622 // no output 2623 } 2624 2625 void 2626 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2627 // no udpate 2628 } 2629 2630 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2631 return 0; 2632 } 2633 2634 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2635 Bytecodes::Code bc) { 2636 // no update 2637 } 2638 2639 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2640 int trap_state) { 2641 jio_snprintf(buf, buflen, "#%d", trap_state); 2642 return buf; 2643 } 2644 2645 #endif // COMPILER2_OR_JVMCI